Data-Efficient Computational Pathology Platform for Faster and Cheaper Breast Cancer Subtype Identifications (Preprint)

BACKGROUND Breast cancer subtyping is a crucial step in determining therapeutic options, but the molecular examination based on immunohistochemical (IHC) staining is expensive and time-consuming. Deep learning model opens up the possibility to predict the subtypes based on the morphological information from hematoxylin and eosin (H&E) staining, a much cheaper and faster alternative. However, training such predictive models conventionally requires a massive number of histology images that leads to high costs and used to be infeasible for single institute. OBJECTIVE We aim to to develop a data-efficient computational pathology platform, 3DHistoNet, that is capable to learn from a relatively small number of histology images for accurate predictions of breast cancer subtypes. METHODS We retrospectively examined 420 cases of primary breast carcinoma patients diagnosed between 2018 and 2020 at the Department of Pathology, National Cancer Center (South Korea) pathology slides of the breast carcinoma patients were prepared according to the standard protocols. Age, gender, histologic grade and hormone receptor (ER, PR) status/HER2 status/Ki-67 index were evaluated by reviewing medical charts and pathological records. RESULTS The area under the receiver operating characteristic curve (AUC) and decision curve were analyzed to evaluate the performance of our 3DhistoNet platform for predicting the ER, PR, AR, HER2 and Ki67 subtype biomarkers with a 5-fold cross validation. We demonstrate 3DHistoNet can predict all clinically important biomarkers (ER, PR, AR, HER2 and Ki67) with performance exceeding the conventional multiple-instance-learning models by a considerable margin (AUC: 0.75-0.91 vs 0.67-0.8). We further show that our novel histology scanning approach can make up the limitation of insufficient training dataset without any additional cost. Finally, 3DHistoNet offers an additional capability to generate attention maps which reveal correlations between histologic features and biomarker expressions. CONCLUSIONS Our platform, with its high prediction capability and versatility, is an appealing tool as an effective prediction tool for the breast cancer subtype biomarkers. Its development would encourage morphology-based diagnosis which is faster and less error-prone compared to the protein quantification method based on IHC staining.